← Frontier Inference Margins · all research reports

Research artifact — each page states its own provenance class in its header: verbatim originals carry SHA-256 stamps; adopted-findings summaries and reconstructions say so explicitly. Conclusions are synthesized (and where needed corrected) in the main report; each page carries its own run date.

Preset grounding pack — GPT-5.6 Pro consultation, 2026-07-10

Run 28m22s · gpt-5-6-pro (Pro-verified, completion_path: api, Ashitaorbis project) · conversation Companion design consultation: consultation-2026-07-10-council-design.md (4-persona council, Opus synthesis), whose ship-list adjudications governed which presets exist. Note: the first run of this consultation completed but expired from the MCP retention window during a usage-limit pause; this is the re-run. The full 192-row JSON/Markdown pack lives in the conversation's sandbox files (retrieve from the conversation if ever needed); the summary below carries every decision the site adopted.

Tariff verifications (all first-party, as of 2026-07-10)

DeepSeek V4-Flash: 284B total / 13B active; $0.14 uncached / $0.0028 cached / $0.28 output; 1M context; 384K max output; concurrency 2,500. → site preset CONFIRMED.
GLM-4.7: 355B-A32B (BF16 + FP8 releases, MTP recipes); $0.60 / $0.11 / $2.20; cached-input storage temporarily free. → CONFIRMED.
GPT-5.6 Terra/Luna: $2.50/$0.25/$15 and $1/$0.10/$6 standard short-context; cache writes 1.25× uncached; Batch/Flex = 50%; >272K whole-request repricing. → CONFIRMED.
Gemini 3.5 Flash: $1.50 / $0.15 / $9 + $1/Mtok-hour cache storage; Batch $0.75/$0.075/$4.50; 1,048,576-token input / 65,536 output; no parameter disclosure. → CONFIRMED.

Locked perspective decisions (adopted)

xAI cash-marginal: $0.60/GPU-hr at the dive's operating point → ~$0.67/M output (site lands ~92% margin on Grok — matches).
xAI opportunity-cost: $5.27/GPU-hr (Anthropic contract: ~325k named NVIDIA GPUs + CPUs/storage/networking at $1.25B/mo — a bundled external opportunity value, not pure accelerator cost) → ~$5.86/M output in its coarser replica model; the site's engine lands ~36% blended vs the dive's ~29% (model-shape divergence documented in the preset note).
China public-cloud on-demand: per-SKU multipliers vs IDC-annual defaults — H20 6.24×, H800 6.07× (monthly HCC, not clean on-demand), 910C-proxy 6.15×; scalar fallback 6.15×; enterprise on-demand utilization 0.35. → site ADOPTED (was 3.5×/45 pre-pack).
Ant H20 SLO replay: representative tier = Pro (675 tok/s, TPOT<50ms) — the middle SLO and the page's neutral anchor; effective MFU 0.16875 ≈ the site's 17% fit; utilization = 1.0 as the per-occupied-chip convention (not a fleet-occupancy claim). → site ADOPTED (util 60→100).

Import semantics (recorded for future preset work)

null = unknown-and-unset, never zero; MFU-override and interactivity-fallback are mutually exclusive; batchShare = published asynchronous batch tariffs only, never internal continuous batching; utilization=1.0 on measured operating points is a convention, not occupancy. Terra/Luna/Gemini-Flash are TARIFF SCENARIOS — size/topology unknown; only tariffs and endpoint metadata carry DISCLOSED.